Physics-aware learning and domain-specific loss design in ophthalmology.

The human cataract, a developing opacification of the human eye lens, currently constitutes the world's most frequent cause for blindness. As a result, cataract surgery has become the most frequently performed ophthalmic surgery in the world. By removing the human lens and replacing it with an artificial intraocular lens (IOL), the optical system of the eye is restored. In order to receive a good refractive result, the IOL specifications, especially the refractive power, have to be determined precisely prior to surgery. In the last years, there has been a body of work to perform this prediction by using biometric information extracted from OCT imaging data, recently also by machine learning (ML) methods. Approaches so far consider only biometric information or physical modelling, but provide no effective combination, while often also neglecting IOL geometry. Additionally, ML on small data sets without sufficient domain coverage can be challenging. To solve these issues, we propose OpticNet, a novel optical refraction network based on an unsupervised, domain-specific loss function that explicitly incorporates physical information into the network. By providing a precise and differentiable light propagation eye model, physical gradients following the eye optics are backpropagated into the network. We further propose a new transfer learning procedure, which allows the unsupervised pre-training on the optical model and fine-tuning of the network on small amounts of surgical patient data. We show that our method outperforms the current state of the art on five OCT-image based data sets, provides better domain coverage within its predictions, and achieves better physical consistency.

Automated Retinal Image Analysis for Evaluation of Focal Hyperpigmentary Changes in Intermediate Age-Related Macular Degeneration.

PURPOSE: To develop and evaluate a software tool for automated detection of focal hyperpigmentary changes (FHC) in eyes with intermediate age-related macular degeneration (AMD). METHODS: Color fundus (CFP) and autofluorescence (AF) photographs of 33 eyes with FHC of 28 AMD patients (mean age 71 years) from the prospective longitudinal natural history MODIAMD-study were included. Fully automated to semiautomated registration of baseline to corresponding follow-up images was evaluated. Following the manual circumscription of individual FHC (four different readings by two readers), a machine-learning algorithm was evaluated for automatic FHC detection. RESULTS: The overall pixel distance error for the semiautomated (CFP follow-up to CFP baseline: median 5.7; CFP to AF images from the same visit: median 6.5) was larger as compared for the automated image registration (4.5 and 5.7; P < 0.001 and P < 0.001). The total number of manually circumscribed objects and the corresponding total size varied between 637 to 1163 and 520,848 pixels to 924,860 pixels, respectively. Performance of the learning algorithms showed a sensitivity of 96% at a specificity level of 98% using information from both CFP and AF images and defining small areas of FHC ("speckle appearance") as "neutral." CONCLUSIONS: FHC as a high-risk feature for progression of AMD to late stages can be automatically assessed at different time points with similar sensitivity and specificity as compared to manual outlining. Upon further development of the research prototype, this approach may be useful both in natural history and interventional large-scale studies for a more refined classification and risk assessment of eyes with intermediate AMD. TRANSLATIONAL RELEVANCE: Automated FHC detection opens the door for a more refined and detailed classification and risk assessment of eyes with intermediate AMD in both natural history and future interventional studies.

Crystal structure of (eth-oxy-ethyl-idene)di-methyl-aza-nium ethyl sulfate.

In the title salt, C6H14NO(+)·C2H5SO4 (-), the C-N bond lengths in the cation are 1.2981 (14), 1.4658 (14) and 1.4707 (15) Å, indicating double- and single-bond character, respectively. The C-O bond length of 1.3157 (13) Šshows double-bond character, indicating charge delocalization within the NCO plane of the iminium ion. In the crystal, C-H⋯O hydrogen bonds between H atoms of the cations and O atoms of neighbouring ethyl sulfate anions are present, generating a three-dimensional network.

Crystal structure of (1-eth-oxy-ethyl-idene)di-methyl-aza-nium tetra-phenyl-borate.

In the cation of the title salt, C6H14NO(+)·C24H20B(-), the C-N bond lengths are 1.297 (2), 1.464 (2) and 1.468 (2) Å, indicating double- and single-bond character, respectively. The C-O bond length of 1.309 (2) Šshows double-bond character, pointing towards charge delocalization within the NCO plane of the iminium ion. In the crystal, C-H⋯π inter-actions between the iminium H atoms and the phenyl C atoms of the anion are present. The phenyl rings form aromatic pockets, in which the iminium ions are embedded.

(But-oxy-methyl-idene)di-methyl-aza-nium tetra-phenyl-borate aceto-nitrile monosolvate.

In the title solvated salt, C7H16NO(+)·C24H20B(-)·C2H3N, the C-N bond lengths in the cation are 1.2831 (19), 1.467 (2) and 1.465 (2) Å, indicating double- and single-bond character, respectively. The C-O bond length of 1.2950 (18) Šshows a double-bond character, pointing towards charge delocalization within the NCO plane of the iminium ion. The two C atoms of the n-butyl group are disordered over the two sites, with refined occupancy ratios of 0.890 (5):0.110 (5) and 0.888 (4):0.112 (4). In the crystal, C-H⋯π inter-actions occur between the methine H atom, H atoms of the -N(CH3)2 and -CH2 groups of the cation, and two of the phenyl rings of the tetra-phenyl-borate anion. The latter inter-action forms an aromatic pocket in which the cation is embedded. Thus, a two-dimensional pattern is created in the ac plane.

3-Methyl-4,5-di-hydro-oxazolium tetra-phenyl-borate.

In the cation of the title salt, C4H8NO(+)·C24H20B(-), the C-N bond lengths are 1.272 (2), 1.4557 (19) and 1.4638 (19) Å, indicating double- and single-bond character, respectively. The C-O bond length of 1.3098 (19) Šshows that double-bond character and charge delocalization occurs within the NCO plane of the cation. In the crystal, a C-H⋯π inter-action is present between the methyl-ene H atom of the cation and one phenyl ring of the tetra-phenyl-borate ion. The latter forms an aromatic pocket in which the cation is embedded.

(Meth-oxy-methyl-idene)di-methyl-aza-nium tetra-phenyl-borate aceto-nitrile monosolvate.

In the cation of the title salt, C4H10NO(+)·C24H20B(-)·C2H3N, the C-N bond lengths are 1.2864 (16), 1.4651 (17) and 1.4686 (16) Å, indicating double- and single-bond character, respectively. The C-O bond length of 1.2978 (15) Šshows double-bond character, pointing towards charge delocalization within the NCO plane of the iminium ion. C-H⋯π inter-actions are present between the methine H atom and two of the phenyl rings of the tetra-phenyl-borate ion. The latter forms an aromatic pocket in which the cation is embedded. The iminium ion is further connected through a C-H⋯N hydrogen bond to the aceto-nitrile mol-ecule. This leads to the formation of a two-dimensional supramolecular pattern along the bc plane.

Choosing the optimal wall shear parameter for the prediction of plaque location-A patient-specific computational study in human left coronary arteries.

OBJECTIVE: While the correlation of atherosclerotic plaque locations with local wall shear stress magnitude has been evaluated previously by other investigators in both right (RCA) and left coronary arteries (LCA), the relative performance of average wall shear stress (AWSS), average wall shear stress gradient (AWSSG), oscillatory shear index (OSI) and relative residence time (RRT) as indicators of potential atherosclerotic plaque locations has not been studied for the LCA. Here we determine the performance of said wall shear parameters in the LCA for the prediction of plaque development locations and compare these results to those previously found in the RCA. METHODS: We obtained 30 patient-specific geometries (mean age 67.1 (± 9.2) years, all with stable angina) of the LCA using dual-source computed tomography and virtually removed any plaque present. We then performed computational fluid dynamics simulations to calculate the wall shear parameters. RESULTS: For the 96 total plaques, AWSS had a higher sensitivity for the prediction of plaque locations (86 ± 25%) than AWSSG (65 ± 37%, p<0.05), OSI (67 ± 32%, p<0.01) or RRT (48 ± 38%, p<0.001). RRT had a higher PPV (49 ± 36%) than AWSS (31 ± 20%, p<0.05) or AWSSG (16 ± 12%, p<0.001). Segment 5 of the LCA presented with overall low values for sensitivity and PPV. Parameter performance in the remainder of the LCA was comparable to that in the RCA. CONCLUSIONS: AWSS features remarkably high sensitivity, but does not reach the PPV of RRT. This may indicate that while low wall shear stress is necessary for plaque formation, its presence alone is not sufficient to predict future plaque locations. Time dependent factors have to be taken into account as well.

Computer-aided detection of ground glass nodules in thoracic CT images using shape, intensity and context features.

Ground glass nodules (GGNs) occur less frequent in computed tomography (CT) scans than solid nodules but have a much higher chance of being malignant. Accurate detection of these nodules is therefore highly important. A complete system for computer-aided detection of GGNs is presented consisting of initial segmentation steps, candidate detection, feature extraction and a two-stage classification process. A rich set of intensity, shape and context features is constructed to describe the appearance of GGN candidates. We apply a two-stage classification approach using a linear discriminant classifier and a GentleBoost classifier to efficiently classify candidate regions. The system is trained and independently tested on 140 scans that contained one or more GGNs from around 10,000 scans obtained in a lung cancer screening trial. The system shows a high sensitivity of 73% at only one false positive per scan.

Computed high concentrations of low-density lipoprotein correlate with plaque locations in human coronary arteries.

Subendothelial accumulation of low-density lipoprotein (LDL) in arterial walls is an initiator of atherosclerotic plaque formation. We report here on the correlation between healthy state subendothelial LDL concentration distribution and sites of subsequent plaque formation in coronary arteries of patients with coronary artery disease (CAD). We acquired left (LCA) and right coronary artery (RCA) and atherosclerotic plaque geometries of 60 patients with CAD using dual-source computed tomography angiography. After virtually removing all plaques to obtain an approximation of the arteries' healthy state, we calculated LDL concentration in the artery walls as a function of local lumen-side shear stress. We found that maximum subendothelial LDL concentrations at plaque locations were, on average, 45% (RCA) and 187% (LCA) higher than the respective average subendothelial concentration. Our results demonstrate that locally elevated subendothelial LDL concentration correlates with subsequent plaque formation at the same location.

Choosing the optimal wall shear parameter for the prediction of plaque location-A patient-specific computational study in human right coronary arteries.

BACKGROUND: Average wall shear-stress (AWSS), average wall shear-stress gradient (AWSSG), oscillatory shear index (OSI) and relative residence time (RRT) are believed to predict areas vulnerable to plaque formation in the coronary arteries. Our aim was to analyze the correlation of these parameters in patients' vessels before the onset of atherosclerosis to the specific plaque sites thereafter, and to compare the parameters' sensitivity and positive predictive value. METHODS: We obtained 30 patient-specific geometries (mean age 67.1 (+ or - 9.2) years, all with stable angina) of the right coronary artery (RCA) using dual-source computed tomography (CT) and virtually removed any plaque present. We then performed computational fluid dynamics (CFD) simulations to calculate the wall shear parameters. RESULTS: For the 120 total plaques, AWSS had on average a higher sensitivity for the prediction of plaque locations (72 + or - 25%) than AWSSG (68 + or - 36%), OSI (60 + or - 30%, p<0.05), and RRT (69 + or - 59%); while OSI had a higher positive predict value (PPV) (68 + or - 34%) than AWSS (47 + or - 27%, p<0.001), AWSSG (37 + or - 23, p<0.001) and RRT (59 + or - 34%). A significant difference was also found between AWSSG and RRT (p<0.01) concerning PPV. CONCLUSIONS: OSI and RRT are the optimal parameters when the number of false positives is to be minimized. AWSS accurately identifies the largest number of plaques, but produces more false positives than OSI and RRT.

Effect of reader experience on variability, evaluation time and accuracy of coronary plaque detection with computed tomography coronary angiography.

OBJECTIVE: To assess the effect of reader experience on variability, evaluation time and accuracy in the detection of coronary artery plaques with computed tomography coronary angiography (CTCA). METHODS: Three independent, blinded readers with three different experience levels twice labelled 50 retrospectively electrocardiography (ECG)-gated contrast-enhanced dual-source CTCA data sets (15 female, age 67.3 +/- 10.4 years, range 46-86 years) indicating the presence or absence of coronary plaques. The evaluation times for the readings were recorded. Intra- and interobserver variability expressed as kappa statistics and sensitivity, specificity, and negative and positive predictive values were calculated for plaque detection, with a consensus reading of the three readers taken as the standard of reference. A bootstrap method was applied in the statistical analysis to account for clustering. RESULTS: Significant correlations were found between reader experience and, respectively, evaluation times (r = -0.59, p < 0.05) and intraobserver variability (r = 0.73, p < 0.05). The evaluation time significantly differed among the readers (p < 0.05). The observer variability for plaque detection, compared with the consensus, varied between kappa = 0.582 and kappa = 0.802. Variability of plaque detection was significantly smaller (p < 0.05) and more accurate (p < 0.05) for the most experienced reader. CONCLUSION: Reader experience significantly correlated with observer variability, evaluation time and accuracy of coronary plaque detection at CTCA.

Guided review by frequent itemset mining: additional evidence for plaque detection.

PURPOSE: A guided review process to support manual coronary plaque detection in computed tomography coronary angiography (CTCA) data sets is proposed. The method learns the spatial plaque distribution patterns by using the frequent itemset mining algorithm and uses this knowledge to predict potentially missed plaques during detection. MATERIALS AND METHODS: Plaque distribution patterns from 252 manually labeled patients who underwent CTCA were included. For various cross-validations a labeling with missing plaques was created from the initial manual ground truth labeling. Frequent itemset mining was used to learn the spatial plaque distribution patterns in form of association rules from a training set. These rules were then applied on a testing set to search for segments in the coronary tree showing evidence of containing unlabeled plaques. The segments with potentially missed plaques were finally reviewed for the existence of plaques. The proposed guided review was compared to a weighted random approach that considered only the probability of occurrence for a plaque in a specific segment and not its spatial correlation to other plaques. RESULTS: Guided review by frequent itemset mining performed significantly better (p < 0.001) than the reference weighted random approach in predicting coronary segments with initially missed plaques. Up to 47% of the initially removed plaques were refound by only reviewing 4.4% of all possible segments. CONCLUSIONS: The spatial distribution patterns of atherosclerosis in coronary arteries can be used to predict potentially missed plaques by a guided review with frequent itemset mining. It shows potential to reduce the intra- and inter-observer variability.

Prediction rules for the detection of coronary artery plaques: evidence from cardiac CT.

OBJECTIVES: To evaluate spatial plaque distribution patterns in coronary arteries based on computed tomography coronary angiography data sets and to express the learned patterns in prediction rules. An application is proposed to use these prediction rules for the detection of initially missed plaques. MATERIAL AND METHODS: Two hundred fifty two consecutive patients with chronic coronary artery disease underwent contrast-enhanced dual-source computed tomography coronary angiography for clinical indications. Coronary artery plaques were manually labeled on a 16-segment coronary model and their position (ie, segments and bifurcations) and composition (ie, calcified, mixed, or noncalcified) were noted. The frequent itemset mining algorithm was used to statistically search for plaque distribution patterns. The patterns were expressed as prediction rules: given plaques at certain locations as conditions, a prediction rule gave evidence--with a certain confidence value--for a plaque at another location within the coronary artery tree. Prediction rules with the highest confidence values were evaluated and described. Furthermore, to improve manual plaque detection, all prediction rules were applied on the patient data to search for segments with potentially missed plaques. These segments were then reviewed in a second, guided reading for the existence of plaques. The same number of segments was also determined by a weighted random approach to evaluate the quality of prediction resulting from frequent itemset mining. RESULTS: In 200 of 252 (79.4%) patients, at least one coronary plaque (range, 1-22 plaques) was found. In total 1229 plaques (990 calcified, 80.6%; 227 mixed, 18.5%; 12 noncalcified, 1%) distributed, over 916 coronary segments and 507 vessels were manually labeled. Four plaque distribution patterns were identified: 20.6% of the patients had no plaques at all; 31.7% had plaques in the left coronary artery tree; 46.4% had plaques both in left and right coronary arteries, whereas 1.2% of the patients had plaques solely in the right coronary artery (RCA). General rules were found predicting plaques in the left anterior descending artery (LAD), given plaques in segments of the RCA or in the left main artery. Further general rules predicted plaques in the LAD, given plaques in the circumflex artery. In the guided review, the segment selection based on the prediction rules from frequent itemset mining performed significantly better (P < 0.001) than the weighted random approach by revealing 48 initially missed plaques. CONCLUSIONS: This study demonstrates spatial plaque distribution patterns in coronary arteries as determined with cardiac CT. Use of the frequent itemset mining algorithm yielded rules that predicted plaques at certain sites given plaques at other sites of the coronary artery tree. Use of these prediction rules improved the manual labeling of coronary plaques as initially missed plaques could be predicted with the guided review.

Patient-specific three-dimensional simulation of LDL accumulation in a human left coronary artery in its healthy and atherosclerotic states.

We calculate low-density lipoprotein (LDL) transport from blood into arterial walls in a three-dimensional, patient-specific model of a human left coronary artery. The in vivo anatomy data are obtained from computed tomography images of a patient with coronary artery disease. Models of the artery anatomy in its healthy and diseased states are derived after segmentation of the vessel lumen, with and without the detected plaque, respectively. Spatial shear stress distribution at the endothelium is determined through the reconstruction of the arterial blood flow field using computational fluid dynamics. The arterial endothelium is represented by a shear stress-dependent, three-pore model, taking into account blood plasma and LDL passage through normal junctions, leaky junctions, and the vesicular pathway. Intraluminal pressures of 70 and 120 mmHg are employed as the normal and hypertensive operating pressures, respectively. By applying our model to both the healthy and diseased states, we show that the location of the plaque in the diseased state corresponds to one of the two sites with predicted high-LDL concentration in the healthy state. We further show that, in the diseased state, the site with high-LDL concentration has shifted distal to the plaque, which is in agreement with the clinical observation that plaques generally grow in the downstream direction. We also demonstrate that hypertension leads to increased number of regions with high-LDL concentration, elucidating one of the ways in which hypertension may promote atherosclerosis.

ACCURATUM: improved calcium volume scoring using a mesh-based algorithm--a phantom study.

To overcome the limitations of the classical volume scoring method for quantifying coronary calcifications, including accuracy, variability between examinations, and dependency on plaque density and acquisition parameters, a mesh-based volume measurement method has been developed. It was evaluated and compared with the classical volume scoring method for accuracy, i.e., the normalized volume (measured volume/ground-truthed volume), and for variability between examinations (standard deviation of accuracy). A cardiac computed-tomography (CT) phantom containing various cylindrical calcifications was scanned using different tube voltages and reconstruction kernels, at various positions and orientations on the CT table and using different slice thicknesses. Mean accuracy for all plaques was significantly higher (p < 0.0001) for the proposed method (1.220 +/- 0.507) than for the classical volume score (1.896 +/- 1.095). In contrast to the classical volume score, plaque density (p = 0.84), reconstruction kernel (p = 0.19), and tube voltage (p = 0.27) had no impact on the accuracy of the developed method. In conclusion, the method presented herein is more accurate than classical calcium scoring and is less dependent on tube voltage, reconstruction kernel, and plaque density.

Automatic detection of calcified coronary plaques in computed tomography data sets.

The detection of calcified plaques is an essential step in the assessment of coronary heart diseases. However, manual plaque segmentation is subjected to intra- and inter-observer variability. We present a novel framework for the automatic detection of calcified coronary plaques in Computed Tomography images. In contrast to the state-of-the-art, both the native and the angio data sets are included to gain additional information about each plaque for its detection and subsequent assessment. The framework was successfully tested on 127 patients where 85.5% of the calcified and 96% of the obstructive plaques have been detected.

Identification of atherosclerotic lesion-prone sites through patient-specific simulation of low-density lipoprotein accumulation.

We present a patient-specific model of low-density lipoprotein (LDL) transport from blood into arterial walls. To this end, the arterial endothelium is represented by a shear-stress dependent three-pore model taking into account blood plasma and LDL passage through the vesicular pathway, normal junctions and leaky junctions. We virtually remove atherosclerotic plaque from an in-vivo left coronary artery computed tomography (CT) dataset to obtain an approximation of the artery anatomy in its healthy state. By applying our model, we show that the location of the plaque in the diseased state corresponds to one of the two sites with predicted high LDL concentration in the healthy state. We further show that in the diseased state, the site with high LDL concentration has shifted distally, which is in agreement with the clinical observation that plaques generally grow in downstream direction.